This policy is for Chromium developers using AI tools to help write code. It is intended to outline expectations around the use of such tools.
Authors must self-review and understand all code and documentation updates (with or without AI tooling) before sending them for review to ensure the correctness, design, security properties, and style meet the standards of the project. Authors should be able to answer questions reviewers have about the changes. Beyond code quality, Chromium has a strict 2-committer code review requirement and when the author is a committer they are considered one of the two human reviewers. Any account that sends for review CLs which are not actually understood by the human behind the account is at risk of losing their committer status. Further violations after being warned may result in the account being banned from the system.
To aid reviewers, authors should flag areas that they are not confident about that had AI assistance. This can be done in code review comments, the CL description, or in code comments. There is a precedent for separating automatically-generated code from manual edits with different patchsets (e.g. patchset 1 has automatic changes and the reproduction instructions and patchset 2+ have manual edits) along with steps to reproduce the automated parts.
Authors must attest that the code they submit is their original creation, regardless of whether AI tooling was used.
Authors may explain in the CL description or the code base itself how AI tools were used to produce the CL.
Examples:
Additional examples for gemini-cli can be added to //agents/prompts/eval, which will serve as eval cases for improvements to common system prompts.
See go/chrome-internal-ai-policy for additional requirements.